I guess someone has to be the negative one: I can't help feeling it's route to the correct face looks entirely accidental (and I don't mean that in a good way)... I'm sure it's "learned" some methods, but they don't look that efficient, reliable, purposeful or controlled. In a more noisy and dynamic environment I'd expect them to fail. Granted is possible these could be more due to training conditions than an inherent limitation of the underlying model.
It looks that way because they're moving rapidly from one face configuration to another. But there's no way that's happening by random. I would guess that even just holding the cube constant in a dynamic grip is quite difficult.
I agree it looks sloppy, but that doesn’t mean it isn’t reliable. All it has to do in any given moment is make progress towards the goal of having the cube in the proper orientation, on average. It may be that it can do that very reliably even with noisy inputs and outputs.
